Evolved Robots Are Born to Run and Refuse to Die

Northwestern Engineering researchers have developed the first modular robots with athletic intelligence. They can be combined and recombined in the wild, recover from injury, and keep moving no matter what’s thrown at them.

Called “legged metamachines,” the creations are made from autonomous, Lego-like modules that snap together into an endless number of configurations. Each module by itself is a complete robot with its own motor, battery, and computer. Alone, a module can roll, turn and jump. But the real agility and indestructibility emerges when the modules combine. 

The study was published March 6 in the Proceedings of the National Academy of Sciences.

To design the most effective combinations, the engineers used AI to evolve novel body configurations. Instead of sticking with standard dog- or human-like designs, the AI churned out strange new “species” of machines that no human engineer would have conceived. When connected to other modules, the metamachines undulate like seals, bound like lizards, or spring like kangaroos.

The robots also can flip themselves upright when turned over, hop over obstacles, and perform acrobatics like spinning in air. Because a metamachine is essentially a robot made up of other robots, it can resist catastrophic damage. Broken parts don’t become dead weight; they keep rolling, crawling, and rejoin the team. 

By combining physical modularity with AI-driven design, the researchers have opened the door to a new class of robots that don’t just survive the real world—they adapt to it. These machines point toward a future where robots are less like fragile, pre-designed tools and more like resilient, evolving lifeforms.

“These are the first robots to set foot outdoors after evolving inside of a computer,” said Northwestern’s Sam Kriegman, who led the study. “They are rapidly assembled and then quite literally hit the ground running. They can move freely in the wild and easily recover from major injuries that would be fatal to every other wild robot. If flipped upside down, they instinctively bring themselves upright and continue their journey. They can survive being chopped in half or cut up into many pieces. When separated, every module within the metamachine can become an individual agent.”

An expert in biorobotics and AI, Kriegman is an assistant professor of computer science, mechanical engineering, and chemical and biological engineering at the McCormick School of Engineering, where he is a member of the Center for Robotics and Biosystems (CRB). The study’s co-first authors are Chen Yu, David Matthews, and Jingxian Wang, who are all PhD students in the CRB.

Evolution accelerated by computers

 While today’s robots can be fast and agile, their body shapes are often fixed and rigid. Most robots cannot adapt to new tasks, environments, or physical damage. If a robotic dog breaks a leg, for example, it’s basically useless. To escape those limitations, Kriegman’s team turned to AI—not to copy familiar designs but to evolve something entirely new.

Kriegman and his team started with an evolutionary algorithm that mimics natural selection. As a starting point, the team gave the algorithm the building blocks for the robot.

These building blocks are half-meter-long modular legs, which look like a pair of sticks joined by a central sphere.

“Inside the sphere, the robot has everything it needs to survive: a ‘nervous system,’ a ‘metabolism’ and ‘muscle,’” Kriegman said. “By that, I mean a circuit board, a battery, and a motor. The modules are mechanically simple. They can only rotate around a single axis, but they are surprisingly athletic and smart.”

Then, Kriegman and his team gave the algorithm a goal: Design a robot with efficient, versatile movement. By mixing and matching the modules in different combinations, the algorithm generated new body types. It then simulated each design, keeping the best performers and discarding the weak. It also iteratively “bred” new designs by combining or mutating them. Depending on the robot’s body, modular legs became legs, spines, or tails.

“We simulated the Darwinian process of mutation and selection within a virtual, physical environment,” Kriegman said. “This is survival of the fittest—accelerated by computers and made real by athletic modular building blocks.”

Traversing rugged terrain

To test the designs, Kriegman and his team assembled the best three-, four-, and five-legged designs found by evolution. In outdoor tests, the metamachines ran across rough terrain, including gravel, grass, tree roots, leaves, sand, mud, and uneven bricks. They jumped, spun, and righted themselves when flipped—all without complicated setup or retraining. 

Unlike traditional robots that fail when a single part breaks, these machines can adapt, recover, and survive. Even when a leg breaks off, the metamachine remains resilient. The modules adapt to a missing leg and keep moving. The missing leg, too, can roll home and rejoin its team.

“It can sense its surroundings, move from place to place, compute, and learn,” Kriegman said. “Metamachines can be rapidly assembled, repaired, redesigned, and recombined. Once assembled, they immediately move themselves across a wide array of unstructured environments.”

The new study builds off previous work from Kriegman’s lab, in which his team designed the first AI algorithm to intelligently design robots from scratch. By compressing billions of years of evolution into mere seconds, the algorithm successfully designed a small, flexible, walking robot in mere seconds. While those robots could not do more than walk across a table, they proved that AI can instantly evolve working robots.

 “Our previously evolved robots couldn’t sense their own bodies or coordinate themselves,” Kriegman said. “But they still taught us a lot about how evolution works and how to distill those lessons into useful technologies. Evolution can reveal new designs that are different from or even beyond what humans were previously capable of imagining. So, we really wanted to study how and why it works. The best way—or at least the most fun way—is to evolve structures in realistic conditions.”